Face Mask Detection Based on Improved YOLOv8

Abstract

The detection of face mask wear is one of the essential measures to prevent the spread of infectious diseases in public places. In order to balance the is-sues of inference speed and performance of target detection models on em-bedded devices, this paper proposes a face mask detection based on the im-proved YOLOv8 algorithm, YOLOv8n-SLIM-DYHEAD. By improving the YOLOv8n algorithm, the balance between detection time and accuracy is-sues is achieved. The Mosaic data augmentation method is used to increase the detection targets of various sizes, enrich the sample dataset of masks of various scales. On the neck network, the Slim-neck structure is used to fuse features of different sizes extracted by the leading network, reducing the complexity of the model while maintaining accuracy. In the detection layer, DyHead is used to integrate better feature diversity caused by target scale differences and target shape position differences. Experimental results show that the improved algorithm YOLOv8n-SLIM-DYHEAD has increased the mAP @0.5 and mAP @0.5:0.95 of the original YOLOv8n algorithm by 2.1 and 5.5 percentage points, respectively. In addition, the complexity and pa-rameters of the model have remained relatively high, and it can accurately detect the wearing of masks in real-time.